Dream-RSI: AI self-improves by dreaming on past discoveries, cutting costly online trials

Dream-RSI: Recursive Self-Improvement through Evolving Worlds

Dream-RSI turns past exploration logs into a replay simulator, letting an agent evaluate and refine its search policy offline without expensive online rollouts. Tested on algorithm engineering, math optimization, and GPU kernel engineering, it matches or beats discovery quality at lower cost. The paper doesn't specify exact savings, but the idea makes recursive self-improvement more practical.